Task- and Usage-Driven Assessment of Low-Contrast Detectability In Fluoroscopy Using AI-Based Phantom Registration and Human Visibility Probabilities
Abstract
Purpose
Contrast-to-noise ratio (CNR) is a common alternative to human visual scoring to assess low-contrast visibility in fluoroscopy systems. We propose a task- and usage-driven framework that combines AI-based phantom registration, image-derived metrics, and human observer labels to automatize the QC process and connect probabilistically image-derived metrics to human low-contrast detectability.
Methods
A TOR-FG18 phantom was imaged across different systems. A YOLO-based localization model automatically identified reference ROIs to register a template mask, enabling calculation of CNR, local background noise, and apparent resolution from line-pair targets using Droege’s method. Human visibility was annotated by three experienced medical physicist, yielding a dataset of insert-level detection outcomes. Analysis performed on the dataset included both a deterministic evaluation and a probabilistic approach.
Results
Visible- and non-visible-insert CNR distributions are statistically different (Wilcoxon–Mann–Whitney test p<0.01). From ROC analysis CNR results a strong discriminator (AUC ≈ 0.94; 95% CI ≈ 0.94–0.95) with an operating point around CNR ≈ 0.6 (Youden threshold; sensitivity ≈ 0.84; specificity ≈ 0.89), but the residual overlap for intermediate CNR values suggest CNR cannot deterministically separate visibility outcomes. Conditional probability showed that both local noise and apparent resolution affect detection. In the multivariable logistic model, log(CNR) and MTF50 contributed independently to detection: β_logCNR = 3.29 (95% CI 2.75–3.84, p<0.001) and β_MTF50 = 1.07 (95% CI 0.66–1.48, p<0.001). Consistently, predicted log(CNR)-detection curves shifted to lower CNR values for increasing MTF50, indicating improved detectability at comparable CNR for higher apparent resolution.
Conclusion
Automatizing low-contrast assessment in C-arm fluoroscopy is feasible. Detectability appears better represented as a probability conditioned on multiple image-quality metrics rather than as a single deterministic threshold on CNR. Combining automated AI-based phantom registration with stratified probabilistic models (CNR with noise and apparent resolution) provides a reproducible, usage-focused approach that more closely reflects observer performance under realistic clinical acquisition variability.